← All work

Own product · 2026 to now · In development

Warrantor

An accountability layer for AI agents. Every agent action is authorised, approved where it matters, and recorded in evidence an auditor can verify without trusting the system that produced it.

Offlineverification of exported evidence

The problem

AI agents are moving from drafting to doing: refunds, record changes, purchases, messages to customers. Every security team and auditor asks the same question. Who allowed this agent to do that, within what limits, and how do we know the record has not changed since?

Agent activity

Today

  • 14:02Refund €240, order 5521refunds-agentApproved by reviewer
  • 13:47Refund €60, order 5498refunds-agentWithin mandate
  • 13:31Credit note €1,900billing-agentWaiting for approval
  • 13:12Refund €3,400, order 5370refunds-agentBlocked: over mandate
  • 12:58Update supplier IBANvendor-agentWaiting for approval

My role

Architect and engineer

Where it stands

580+
automated tests on the core
GDPR
erasure that keeps the audit trail intact

Stack

  • Python
  • Cryptography
  • SQLAlchemy
  • Typer

What changes

Authority

Without“The model decided.”

With WarrantorEvery action runs under a mandate granted by a named person, with limits.

Oversight

WithoutAgents act first. People find out later.

With WarrantorActions that need sign-off wait for a named approver.

Evidence

WithoutLogs the operator can edit.

With WarrantorA tamper-evident record with a signed receipt per decision, verifiable offline.

Privacy

WithoutPersonal data in an audit trail cannot be erased without breaking it.

With WarrantorGDPR erasure that keeps the trail intact.

How it works

At a high level. A detailed walkthrough is available on a call.

  1. 01Mandate
  2. 02Decide
  3. 03Approve
  4. 04Record
  5. 05Protect
  6. 06Verify
  1. Decision 01

    Check the mandate before the policy, on every action.

    An agent without authority is denied before anyone debates whether the action is a good idea.

  2. Decision 02

    Delegation can only narrow authority.

    An agent can hand a task to another agent, but never more power than it was given.

  3. Decision 03

    Evidence that verifies offline.

    An auditor should not have to trust the vendor, the operator or the agent to check the record.

  4. Decision 04

    Fail closed.

    When a check cannot be completed, the action does not run.